Garry Tan's brain holds 146,646 pages.
Not the brain in his skull (obviously). A second brain that lives on his computer, one his AI agents read before they do anything on his behalf.
In our last issue I gave you the four layers of agent memory, and my rule 1 was "when in doubt, it goes to GBrain." Then my inbox did what my inbox does. Ok Justin, but what IS GBrain? Is it hard to install? Do I really need it for my one little support agent?
Fair questions. So today you're getting the whole story. The founder and the problem that forced him to build it, how it works under the hood, what it costs (almost nothing), how it stacks up against the alternatives, and an honest list of who should NOT install it yet.
The Guy Who Runs Y Combinator Had a Memory Problem
Garry Tan runs Y Combinator. That's the startup school that hatched Airbnb, Stripe, Dropbox, DoorDash, and Reddit.
Think about what that job actually is. Thousands of founders. Endless demo days, board meetings, intros, emails, and DMs. His entire value is knowing who's building what, who invested in whom, what he promised to whom, and what he learned in a meeting 14 months ago.
And his AI agents had the same disease yours do. Smart but forgetful. (He runs his on OpenClaw and Hermes, two popular agent harnesses. A harness is the software body that lets an AI use tools and act on its own instead of just chatting. Hermes is the same one my agents run on.) Every conversation started from a blank page, with the guy who runs Y Combinator re-pasting context like it was the intern's first day. Forever.
You know that pain at YOUR scale. Now multiply it across 24,585 people and 5,339 companies, because that's how many his system tracks today.
So he built his agents a brain. It swallows his meeting notes, emails, tweets, voice calls, and random shower ideas, and files each one as a markdown page. (Markdown is just plain text with a few simple symbols for formatting. Those .md files open in any notes app. If you can type an email, you can read markdown.) Then it wires all the people and companies together into a living map, and 66 automated jobs polish that brain overnight while he sleeps.
On April 5th he open-sourced the whole thing, meaning he published the code free for anyone to use or inspect. That was one day after Karpathy dropped his LLM Wiki gist (a gist is a little public note programmers share). It pulled around 5,000 stars in its first 24 hours on GitHub, the site where programmers share code (stars are GitHub's version of likes), and it sits above 26,300 today. The license is MIT, which is lawyer-speak for "do whatever you want with it, forever, free." So nobody can ever take it away from you.
Garry's own description: GBrain is "a knowledge system, not RAG in a box." (RAG is nerd-speak for the standard trick of letting an AI search your documents and paste what it finds into its answer. Garry's saying GBrain goes well past that.) It's built to make your agent "feel clairvoyant about who you are," and his closing line was "Personal AI becomes possible."
Clairvoyant is a big word. Let's pop the hood and see if he earned it.
How GBrain Works (The Library Tour)
The easiest way to understand GBrain is to walk through it like a library. Stay with me, because each room gets better.
The books. Every memory is a markdown page stored in a git repo. (A git repo is a folder that keeps a dated history of every change ever made inside it. Think Track Changes in Word, but for a whole folder.) One page per person, per company, per project. Each page keeps the compiled truth OVER a line and the dated receipts UNDER it, so you can audit any fact like a bank statement. Out of the box you get 15 page types: person, company, deal, email, tweet, project, note, and so on. And you can change them, because it's YOUR library.
The card catalog. Here's where GBrain leaves a plain folder behind. Every page gets indexed three ways at once. Vector search finds MEANING: the AI files each page by what it's ABOUT, so a question like "that Nashville guy's complaint" finds the right page even when your words don't match its words. Keyword search finds exact names and terms, plain old Ctrl+F style. And the knowledge graph is a relationship map: picture a detective's corkboard with string running from every person to every company, except every string is labeled (works_at, invested_in, founded). GBrain runs all three searches, blends the results, and floats the best pages to the top.
On their published benchmark (their own accuracy test), the right page lands in the top 5 results 97.9% of the time, and the relationship map alone makes the results 31.4 points more precise than meaning-search by itself. It's their corpus and their test, so hold it loosely. But they published exact numbers on a clearly described setup instead of screaming "beats every benchmark!!" like most AI tools do, and I respect that.
The librarian. This is the feature that sold me. GBrain gives you two commands (you can type them yourself, or your agent runs them for you). gbrain search hands you ranked pages, like Google. But gbrain think hands you a written ANSWER with citations to the exact pages it used... plus a list of what's stale, what contradicts what, and what's missing. Garry's pitch line: "Search gives you raw pages. GBrain gives you the answer."
The self-wiring trick. When a page gets written, GBrain spots the people and companies in it and wires up the relationship map using simple find-and-match rules. Basically a very smart Ctrl+F, with zero AI involved in that step. That matters because AI calls are the part that costs money, so your library wires itself every single day without adding a dime to your AI bill. (My favorite engineering choice in the whole build. Boring, cheap, and bulletproof.)
It notices who keeps showing up. Mention someone once and they get a stub page. Three mentions and GBrain goes off to research them on the web without being asked. A meeting or 8+ mentions triggers the full workup. Which is exactly how you'd treat the customer who emails you five times, except it never forgets to do it.
The dream cycle. Overnight, on a schedule, GBrain dedupes records, fixes citations, flags contradictions between pages, scores what matters, and preps tomorrow's work. Garry runs 66 of these jobs. Your agents' brain gets sharper on Tuesday than it was on Monday, and nobody touched it.
The staff. It ships with 43 skills (plain-text playbooks for capturing, importing, researching, and auditing) plus a job queue called Minions that runs durable little workers who survive crashes. And you can feed the brain from anywhere: typed commands, a Zapier zap, an iPhone shortcut, email, a webhook (an automation doorbell that other apps can ring to drop something off)... you can even CALL your brain on the phone and talk to it while driving (Twilio, the service that gives software a phone number, handles the call).
That's the library. Books you own, a card catalog with three indexes, a librarian who already read everything, and a night crew that tidies up while you sleep.
What It Costs and What It Runs On
The software is free. That MIT license means free forever, with no subscription hiding in the bushes.
For a personal brain it runs on PGLite, a tiny database that installs itself and boots in about 2 seconds. (A database is just software that stores records and finds them fast. You never touch this part. It's the engine under the hood.) No server to rent, no database class to take, and it carries you to roughly 50,000 pages, which is years of runway for a normal operator. Outgrow it and you upgrade to Postgres, the industry's heavy-duty database (Supabase will host one for you, and their free tier works fine).
The only real bill is embeddings, which is the technical name for that meaning-math filing step. An AI reads each page once and files it by meaning, and that one-time read is what you pay for. Filing 7,500 pages costs about $4-5, and an active personal brain runs single-digit dollars a month after that.
The install is my favorite part. The repo ships an install guide written for AGENTS, not humans. You paste it into Claude Code, OpenClaw, or Hermes, and your agent installs its own brain in about 30 minutes. After that it plugs into Claude Desktop, Cowork, ChatGPT, Cursor, and friends through MCP. (MCP stands for Model Context Protocol, but all you need to know is that it's the universal plug that lets an AI log into your other tools. It's the same plug my support agent uses to drive our help desk.)
One warning, and I mean it: install from the official repo ONLY, meaning the project's official page on GitHub: github.com/garrytan/gbrain. The week this blew up, squatters parked fake "gbrain" copies on the sites programmers install software from, hoping you'd typo your way into their code. The project's own README (the instruction sheet on its front page) warns you about this. Real repo or nothing.
GBrain vs Everything Else
Fair fight time, because "just use GBrain" is lazy advice and you deserve the honest matchup.
GBrain vs a plain memory folder. The folder is still your correct first move (we covered that build in Build Notes #020). GBrain is that SAME markdown with the card catalog and the night crew bolted on. Upgrade when your agent starts missing things you KNOW are in the folder. You'll feel the line when you cross it.
GBrain vs a vector database. A standalone vector database (that's meaning-math storage with no readable files) is impressive recall you can't read, can't audit, and can't fix with a text editor. GBrain keeps the readable files AND adds vectors on top, plus the graph that bumped precision 31.4 points. You lose nothing and you gain receipts.
GBrain vs ChatGPT's and Claude's built-in memory. Built-in memory is sticky notes in someone else's office. Helpful, but you can't export it, can't audit it, can't share it across tools, and it's hostage if you ever switch. GBrain memory lives in YOUR files. It survives model swaps, works in every tool at once through MCP, and nobody can shut it off. Anyone can rent your model. Nobody can rent your memory.
GBrain vs Mem0, Zep, Letta, and the memory-as-a-service crowd. Those products rent memory infrastructure to software teams building apps where thousands of users each need their own memory. Different animal entirely. If that's you, go look at those. If you're building YOUR business a brain that YOU own, that's the exact reason GBrain exists.
And the honest cons. It's young (version 0.30 as I write this, which is software-speak for early days), and updates still break things between versions. First-class support really means OpenClaw and Hermes; everything else, including Claude, connects through that MCP plug, and setting the plug up is on you. It's built for ONE operator, not a team sharing one brain. There's no managed cloud version, so you're the host. And it won't invent organization for you: if your facts don't fit an existing skill, you write a new skill. It's a workshop tool, not an appliance.
So if you need team-shared memory, enterprise compliance, or zero maintenance... wait, or go with the managed crowd. For a one-person AI operation that wants to own its brain? I haven't found anything close.
Do You Actually Need It?
The adoption ladder, honestly:
- Just chatting with AI? Built-in memory is fine. Skip all of this.
- Running your first agent? Start with a memory/ folder and the 4 memory rules from Build Notes #020. Free, 30 minutes, no excuses.
- Agent starts missing things you KNOW it knows? That's the line. Install GBrain, point it at the same markdown folder, and keep moving.
I crossed that line when our support agents needed every customer's full history at ticket speed. A folder can HOLD 900 customer files no problem. Finding the right three lines across 900 files in two seconds is a retrieval problem, and retrieval is the entire reason GBrain exists.
Your Homework This Week
- No memory folder yet? Build that first: a folder of markdown notes with an index.md as the table of contents and a log.md as the diary. GBrain with nothing to feed it is a library with no books.
- Feeling the misses? Paste github.com/garrytan/gbrain into Claude Code (or OpenClaw, or Hermes) and tell your agent to follow the install guide for agents. Budget 30 minutes.
- Import what you have: your notes folder, your customer files, your project docs. The brain starts empty until you feed it.
- Point your agent's default long-term memory rule at GBrain.
- Let the dream cycle run for one week. Then ask
gbrain thinka question you never directly told it the answer to. The first time it answers WITH receipts, you'll understand why the guy who sees every startup on Earth built this himself instead of buying something.
Continue with Build Notes+
This Monday deep dive is free in full. For complete Thursday builds and supporting files, join Build Notes+.
